How Castform + Neon Beats Frontier Models on Price and Efficiency

Castform and Neon describe a workflow for reinforcement-learning post-training of small open-source models to perform agentic search over company data, claiming that a 4B model trained this way matched GPT-5.6 Sol’s retrieval accuracy at roughly one-hundredth the cost. The approach responds to a shift from one-shot embedding-based RAG systems toward multi-step agentic retrieval, where frontier-model search loops can add substantial latency and expense. Castform generates synthetic questions and ground-truth answers from documents stored in Neon Postgres, trains models to use Lakebase Search’s text, vector, and hybrid retrieval tools, and evaluates them through rewards for finding correct sources, citing them, and answering accurately. The platform also provides training-run observability to inspect reward progress and diagnose issues such as faulty tools or reward hacking. Neon’s autoscaling database infrastructure is presented as useful for the bursty workloads created by many parallel search rollouts, while its branching and time-travel capabilities could support isolated, resettable environments for training agents that modify data as well as retrieve it.

Use this post, company, and trend context to find content marketing opportunities, perform competitive analysis, or address product feature gaps via the Plushcap MCP server or the Plushcap API.